On Mon, 14 Jul 2025 15:43:58 GMT, Raffaello Giulietti <rgiulie...@openjdk.org> wrote:
>> @rgiulietti The convergence of the recurrence is guaranteed if the initial >> estimate is larger than or equal to the exact value, AFAIK no guarantee is >> given when the estimate is smaller than the exact value. > > This is my hunch as well. > So while the use of `nextUp()` and `ceil()` certainly help to achieve an > overestimate, I'm less sure that this is sufficient when using `pow()`. If > the latter has some error bigger than a few ulps, then we are in trouble. @rgiulietti Now that I checked the recurrence better, it seems to me that it should give an overestimate if `rLong` is an underestimate, without modify it. ------------- PR Review Comment: https://git.openjdk.org/jdk/pull/24898#discussion_r2205358875
